Skip to content

Latest commit

 

History

History
182 lines (123 loc) · 7.52 KB

api.md

File metadata and controls

182 lines (123 loc) · 7.52 KB
Raw

MineLand API Documentation

The structure of this document is based on MineDojo Docs, whose clarity and widespread acceptance.

1. First Agent in MineLand

Quick Start

After following the Installation Docs, you can import MineLand.

import mineland

Secondly, an Environment can now be created.

mland = mineland.make(
	task_id="survival_0.01_days",
    agents_count=2,
)

In the code above, task_id refers to the task content executed by the environment. The available options for task_id can be found in Tasks Module. If you do not wish to enable any benchmarks, you can initialize a simulator without benchmark by directly using the mineland.MineLand class, or by setting task_id to playground.

agents_count represents the number of agents initially present in the environment. If you wish to configure the initial settings (like names) of the agents, you can use agents_config. for more details, see agents_config.

MineLand supports connecting to any vanilla 1.19 Minecraft server, meaning that you can use MineLand not only as a machine learning tool but also as a fully functional, interruption-capable, Python-side Minecraft bot library. This has already been applied in some projects. You can pass the parameters server_host and server_port to specify the target server; for more details, see server_host. If these parameters are not provided, a Minecraft server is automatically invoked locally.

Then, let bots connect to the server.

obs = mland.reset()

After mland.reset() has been executed, all bots will connect to the server.

If you haven't enable the PAUSE mode or you connect to a remote server, the time will begin to pass...

Next, we can implement a main loop.

for i in range(5000):
    act = mineland.Action.no_op(len(obs))
    obs, code_info, event, done, task_info = mland.step(action=act)
    if done: break

In the main loop, we need to complete at least 3 parts of the code: getting actions, passing in actions and receiving observations, and checking termination conditions.

First part, getting actions. MineLand's default high-level actions differ from the action spaces of other gym-style libraries. A high-level action consists of a piece of JavaScript code. In this example, we will fetch a RESUME Action (which performs no action). You can also use mineland.Action.chat_op(num_of_agents) to get the bot to print messages in the chat area.

Please note, to implement the "interruption" feature, high-level actions are divided into two modes: RESUME and NEW. The code is executed only under NEW, and any previous code is interrupted. You can check in the observation space whether the previous code has finished executing. For more details, see code_info

Second part, getting observations. Pass the action into the step function and obtain five pieces of information: obs, code_info, event, done, and task_info. Here, obs, code_info, and event are all lists, each equal in length to the number of agents. For instance, obs[0] can be used to view all observation information for the first agent; you might try printing obs[0], code_info[0], and event[0] directly. Meanwhile, done and task_info are not lists, and relate to benchmark information.

Third part, termination conditions. This can be implemented using a simple if done: break statement.

Finally, close the environment by executing the follows.

mland.close()

Observation and Action Spaces

Observation Space is defined in observation.py and observation_utils.js.

Action Space (High-level) is all functions you can invoke in javascript and mineflayer (including pathfinder, etc). You can refer to high_level_action and Mineflayer API Docs.

2. Environment Parameters

Simulator

  • You can refer to: ./mineland/sim/sim.py and ./mineland/tasks/__init__.py

agents_count

  • Type: int
  • Definition: The number of agents in the initial environment.
  • Default value: N/A
  • Example: agents_count=2.

agents_config

  • Type: List[Dict[str, Union[int, str]]],
  • Definition: The configuration of agents.
    • Currently only supports agents' names.
  • Caution: the length of configurations must be equal to agents_count.
  • Default value: [{'name': f'MineflayerBot{i}'} for i in range(kwargs['agents_count'])]
  • Example: above

task_id

  • Type: str
  • Definition: The task you're preparing to activate.
    • The document of tasks is WIP, you can refer to ./mineland/tasks/__init__.py and ./mineland/tasks/description_files/
  • Default value: N/A
  • Example: task_id='survival_2_days',

ticks_per_step

  • Type: int
  • Definition: The number of ticks that pass per single step.
  • Caution: if you want to pause the game while the agent is thinking, you must enable enable_auto_pause.
  • Default value: N/A
  • Example: ticks_per_step=20
    • means waiting 1 second per step

enable_auto_pause

  • Type: bool
  • Definition: Determines whether to enable AUTO PAUSE mode.
    • AUTO PAUSE mode automatically pausees the game when the step function is not running.
  • Default value: False
  • Example: enable_auto_pause=True

enable_sound_system

  • Type: bool
  • Definition: Determines whether to enable Sound System.
  • Caution: Sound System is not complete, but it can still function.
  • Default value: False
  • Example: enable_sound_system=True

server_host

  • Type: str
  • Definition: The host of vanilla 1.19 minecraft server you want to connect.
  • Default value: None
    • Means, MineLand will launch a minecraft server automatically.
  • Example: server_host='localhost'

server_port

  • Type: int
  • Definition: The port of vanilla 1.19 minecraft server you want to connect.
  • Default value: None
  • Example: server_port=25565

headless

  • Type: bool
  • Definition: Determines whether to disable the visual information (RGB frames)
  • Default value: False
  • Example: headless=True

image_size

  • Type: Tuple[int, int]
  • Definition: The size of visual information of agents. (The size of RGB frames)
  • Default value: (144, 256)
  • Example: image_size=(114, 514)

is_printing_server_info

  • Type: bool
  • Definition: Whether to output server information
  • Default value: True
  • Example: is_printing_server_info=False

is_printing_mineflayer_info

  • Type: bool
  • Definition: Whether to output mineflayer information
  • Default value: True
  • Example: is_printing_mineflayer_info=False

3. Observation Space

  • All information about environment and benchmark.
  • You can execute ./scripts/feature_observation_space.py to check observation space.

4. Action Space (High-level)

  • Almost all of the mineflayer API (including pathfinder, collectBlock, etc..)
  • You can use ./scripts/action_space_high_level.py to control the agent by entering some codes (high-level actions).

5. Action Space (Low-level)

  • WIP